Poster Competition

Showcase your research and connect with peers at our annual YinzOR Poster Session, open to all PhD and graduate students in Operations Research, Management Science, Industrial Engineering, and related areas. Present a poster on any topic in your field. Outstanding entries will receive cash prizes!

Accepted Posters

Efficiency of Decentralized Data Sharing Coalitions under Differential Privacy

Diptangshu Sen, Georgia Institute of Technology

Motivated by the rapid push to decentralize data sharing, we study whether large-scale data sharing coalitions can form in a decentralized manner under differential privacy when players have heterogeneous privacy preferences. We provide a comprehensive analysis across multiple privacy-cost regimes corresponding to different attack/observation models in differential privacy. Our analysis shows that full decentralization is "useful" only under limited privacy-cost regimes. Further, it is highly inefficient (the efficiency gap grows polynomially in the number of players) compared to the socially optimal baseline mechanism with full centralization where players have no autonomy, instead a central designer forms coalitions and assigns privacy levels. Surprisingly, we show that a simple intermediate mechanism, which is partially decentralized, recovers most of the nice properties of the baseline mechanism and closes the efficiency gap down to constant factors in all privacy-cost regimes. In summary, our work provides rigorous mathematical foundations for the design of efficient data-sharing ecosystems of the future.

Post-Estimation Adjustments in Data-Driven Decision-Making with Applications in Pricing

Guan Wang, University of Toronto

The predict-then-optimize (PTO) framework is a standard approach in data-driven decision-making, where a decision-maker first estimates an unknown parameter from historical data and then uses this estimate to solve an optimization problem. While widely used for its simplicity and modularity, PTO can lead to suboptimal decisions because the estimation step does not account for the structure of the downstream optimization problem. We study a class of problems where the objective function, evaluated at the PTO decision, is asymmetric with respect to estimation errors. This asymmetry causes the expected outcome to be systematically degraded by noise in the parameter estimate, as the penalty for underestimation differs from that of overestimation. To address this, we develop a data-driven post-estimation adjustment that improves decision quality while preserving the practicality and modularity of PTO. We show that when the objective function satisfies a particular curvature condition, based on the ratio of its third and second derivatives, the adjustment simplifies to a closed-form expression. This condition holds for a broad range of pricing problems, including those with linear, log-linear, and power-law demand models. Under this condition, we establish theoretical guarantees that our adjustment uniformly and asymptotically outperforms standard PTO, and we precisely characterize the resulting improvement. Additionally, we extend our framework to multi-parameter optimization and settings with biased estimators. Numerical experiments demonstrate that our method consistently improves revenue, particularly in small-sample regimes where estimation uncertainty is most pronounced. This makes our approach especially well-suited for pricing new products or in settings with limited historical price variation.

Proactive Inpatient Bed Requests for Emergency Department Admissions

Aniruddhan Ganesaraman, University of North Carolina at Chapel Hill

Emergency department (ED) boarding, the practice of keeping admitted patients in the ED while awaiting inpatient beds, is a primary driver of crowding, poor outcomes, and excess cost. Under current practice, the Transfer Preparation Process begins only after a formal admission decision is made. Initiating bed requests earlier, however, risks having prepared beds remain unoccupied for extended periods. We propose a framework that manages this trade-off directly. A centralized decision-maker periodically determines how many inpatient beds to request based on real-time ED census, patient admission probabilities, and current hospital bed availability. Aggregating requests at the population level rather than the patient level reduces prediction uncertainty and aligns better with operational practice than individualized early requests. The framework is formulated as an infinite-horizon Markov decision process with three cost components: a penalty for boarding patients, a penalty for unoccupied prepared beds, and a convex ordering cost that discourages large request batches. Three heuristic policies are derived: an Markov Decision Process (MDP)-based heuristic (MHP), a newsvendor-based heuristic (NHP), and a deep Q-learning policy (DQN). Policies are evaluated across a wide spectrum of boarding-to-idling cost ratios in a validated high-fidelity discrete-event simulator calibrated to 2019 data from a large academic ED in North Carolina. Under standard conditions, the framework reduces average boarding time by 30–70% and average ED length of stay by 6–15% relative to current practice, with only modest increases in bed idle time. These gains persist when only a fraction of hospital beds are eligible for early requests, confirming robustness to the assumption that beds are interchangeable across patients. Robustness is further confirmed under a pandemic-surge scenario characterized by elevated patient arrival rates. Among the proposed policies, NHP most efficiently balances boarding time reduction against bed idle time, and offers interpretability advantages suited to healthcare settings; DQN achieves the lowest variability in request sizes, producing the smoothest bed request process.

Rounding Weighted Sums of Positive Semidefinite Matrices

Haeseong Yang, University of Pittsburgh

We consider the problem of \emph{rounding a weighted sum of positive semidefinite} (PSD) \emph{matrices}: given $n \times n$ PSD matrices $A_1,\ldots,A_m \succeq 0$ and weights $w \in [0,1]^m$ such that $\sum_{i=1}^m w_i = k \in \mathbb{N}$, the goal is to find a subset $J \subseteq \{1,\ldots,m\}$ of cardinality at most $k$ under which $\sum_{i \in J} A_i \succeq \alpha \sum_{i=1}^m w_iA_i$ and the value of $\alpha \in [0,1]$ is as large as possible. The problem captures an approach that can be applied to round relaxed solutions of various $NP$-hard \emph{spectral combinatorial optimization} problems, including \emph{optimal linear experimental design} and the \emph{graph augmentation} problems. We propose and study \emph{Split Randomized Rounding} (SRR), a randomized rounding algorithm that splits the weighted sum $\sum_{i=1}^m w_iA_i$ into three weighted sums and applies a different rounding procedure (two of which are randomized) to each sum. One of the randomized rounding procedures draws from the sparse sums of PSD matrices literature, while the other is motivated by a matrix Chernoff bound. For $\epsilon > 0$, we show that SRR returns a solution under which we can take $\alpha = 1-\epsilon$ provided that $k = \Omega(r \ln(r) / \epsilon^3)$, where $r \leq n$ is the rank of the weighted sum $\sum_{i =1}^m w_iA_i$. The result implies that SRR can be applied to obtain \emph{$(1-\epsilon)$-approximations} to $NP$-hard spectral combinatorial optimization problems, assuming that $k$ is sufficiently large. We also show that our lower bound on $k$ is tight. Finally, we explore the empirical performance of SRR through computational experiments.

Trustworthy Probabilistic Forecasting of Severe Weather Hazards

Elizabeth Cucuzzella, Carnegie Mellon University

Existing numerical weather prediction systems or AI models can output predictions of, for example, wind speed but they generally do a poor job at predicting extreme and rapidly developing weather events, such as rapidly intensifying tropical cyclones on a 24-hour time scale. The outputted uncertainties are only calibrated on average at best, and there is a lack of diagnostic tools that forecasters and scientists can use to validate predictive distributions. Diagnostic transport maps provide a computationally efficient method for producing interpretable local diagnostics and a mechanism for adjusting predictive densities (PDs) so as to be consistent with observational data. This approach allows us to learn an interpretable probability-probability map from calibration data that identifies how an initial model may be ill-estimated in the input space. Previous work has learned this map through monotonic neural networks for one-dimensional applications, which requires large amounts of training data to perform well. Here we show how a more efficient parametric approach can identify the evolutionary modes for which the National Hurricane Center’s short-term (24-hour) intensity PDs are inaccurate, and we extend the framework to encompass multiple dimensions via conditional copulas. We can use our parametric mapping to reshape the NHC’s initial PD to obtain conditionally calibrated predictive distributions for intensity and track forecasting. Not only does this method provide reliable uncertainties and interpretable results for informed decision making, but it also leads to improved predictive performance over the NHC’s operational forecasts.

Beyond Predicting Responses: Conformal Inference for Latent Distributional Parameters

Minxing Zheng, Carnegie Mellon University

Many prediction problems seek to infer an unobserved, instance-specific parameter that governs the distribution of an observable response, even though the latent parameter is unavailable for both historical and future instances. We develop LatentCP, a prior-free conformal framework that constructs uncertainty sets for latent distributional parameters using only observed context–response pairs and a specified forward model. The method first constructs a conformal prediction set in the observable response space and then retains candidate latent parameters according to the probability their induced response distributions assign to that set. This inversion provides finite-sample marginal coverage without requiring latent calibration labels, a unique inverse mapping, or knowledge of the latent mixing distribution. Because latent-set efficiency depends nonmonotonically on the response-space miscoverage rate, we further introduce a multilevel procedure that aggregates normalized incompatibility scores across several response sets and selects the aggregation distribution using an independent tuning sample. Across synthetic experiments, LatentCP maintains nominal latent coverage under weak forward identification, observational nonidentifiability, and latent heterogeneity and multimodality, where empirical-Bayes, likelihood-based, and proxy-label conformal methods can substantially under-cover. On a California wildfire real dataset, it produces spatially adaptive uncertainty sets for latent fire intensity. Independent tuning improves efficiency, while multilevel aggregation provides additional gains when different response levels contain complementary information, without sacrificing validity

When Waiting Changes the Customer: The Quality–Urgency Tradeoff in Service Systems with Impatient Customers

Bihan Chatterjee, Georgia Institute of Technology

In many service systems, waiting is not merely a source of delay; it can also change the customers or jobs awaiting service. A patient may deteriorate, a support request may become more urgent, a job may lose value, and a customer may eventually abandon the system before receiving service. These effects create a fundamental tradeoff between urgency and prevention: should the server prioritize customers in later stages who are closer to abandonment, or customers in earlier stages before their condition deteriorates? This talk studies a finite-capacity, single-server queue in which customer impatience evolves stochastically through multiple stages while customers remain in the system. This formulation relaxes the common assumption that abandonment occurs after a single exponential patience time, allowing a customer’s impatience and service characteristics to evolve as the customer waits. Each customer begins in an initial stage and may progress through subsequent stages before either completing service or abandoning from the final stage. The service rate, holding cost, and reward from service completion may depend on the customer’s stage, while abandonment incurs a penalty. The objective is to determine which customer stage to serve at each time in order to maximize the long-run average profit. We first consider the setting in which customer stages are fully observable. To obtain structural insight, we analyze two simplified regimes. The first is a congested model in which the system remains full, representing settings in which the system spends most of its time near capacity. In this regime, we obtain a complete characterization of the optimal priority rule for the two-stage model. The second is a light-traffic regime with exogenous arrivals, which reveals when different priority rules can be distinguished asymptotically. Together, these analyses provide conditions under which the server should prioritize urgency (customers closer to abandonment), or quality (customers in earlier stages). We translate these insights into scheduling heuristics and evaluate their performance numerically. Finally, we examine settings in which customer stages are not directly observable. We evaluate simple queue-position-based policies, including FCFS, LCFS, and threshold hybrids, and compare their performance with the fully observable benchmark. We also quantify the long-run average reward gap between fully observable and position-based policies and show that this gap can be closed when the server observes increasingly informative proxy signals.

Fair Supervised Learning Through Constraints on Smooth Nonconvex Unfairness-Measure Surrogates

Zahra Khatti, Lehigh University

A new strategy for fair supervised machine learning (ML) is proposed. Its advantages compared to others are as follows. (a) We introduce a new smooth nonconvex surrogate to approximate the Heaviside functions involved in discontinuous unfairness measures. The surrogate is a tight approximation that ensures the trained prediction models are fair, as opposed to other (e.g., convex) surrogates that can fail to lead to fair prediction models. (b) Rather than rely on regularizers (that lead to optimization problems that are difficult to solve) and corresponding regularization parameters (that can be expensive to tune), we propose a strategy that employs hard constraints so that specific tolerances for unfairness can be enforced. (c)~Our strategy readily allows for constraints on multiple (potentially conflicting) unfairness measures at the same time. Multiple measures can be considered with a regularization approach, but at the cost of having even more difficult training problems and further expense for tuning. By contrast, through hard constraints, our strategy leads to training problems that can be solved tractably through minimal tuning.

Online Prediction Intervals with Prior Knowledge: A Conformal Approach for Time Series

Selina Carter, Carnegie Mellon University

We develop a new method for time series prediction intervals. We assume access to a bank of existing (finite or online) trajectories. We then predict on a new sequence that is streaming online and is right-censored (i.e., we don't know when the trajectory will end). Our prediction method builds upon existing conformal prediction methods that assume a single time series. There are three main outputs: (1) we develop a new algorithm that incorporates the previous bank of sequences to predict S-step-ahead states and prediction bounds; (2) we show analytically that this algorithm reduces the prediction interval width compared to having no prior bank of data, while also maintaining correct theoretical coverage; (3) in a simulation study, we will test the proposed algorithm against baseline techniques. As a use case, we primarily focus on Tokamak plasma dynamics, a challenging problem in nuclear energy research. Other potential use cases are abundant in finance, robotics, and health.

Closing the Uncertainty Gaps in Aircraft Landing Optimization: A Comparative Study of Stochastic, Robust, and Rolling-Horizon Strategies

Bandar Malki, University of Akron

This paper addresses the critical gap between deterministic aircraft landing optimization and the inherently uncertain operational environment of terminal airspace. While the standard Aircraft Landing Problem (ALP) optimizes schedules based on single estimated arrival times and fixed separation matrices, real-world operations face uncertain arrivals, stochastic wake separation requirements, touchdown execution errors, and dynamic traffic streams. We systematically identify five sources of uncertainty and propose minimal-change treatments that preserve the original MILP's efficient binary structure: (1) two-stage sample-average approximation for arrival-time uncertainty, (2) chance-constrained separation for stochastic wake requirements, (3) budgeted robust optimization for touchdown-execution errors, (4) rolling-horizon re-optimization for the open online problem, and (5) distributionally robust separation for unknown noise distributions. Using a Munich airport instance (40 aircraft, 2 runways) and out-of-sample evaluation across 60 scenarios, we demonstrate that separation hedging is remarkably cost-effective: the Gaussian chance constraint reduces separation interventions by 54% with only 2.2% objective increase, while robust buffering achieves 80% intervention reduction at 3.6% cost. Rolling re-optimization delivers the largest single improvement—27% lower realized objective and 41% lower mean delay—but does not hedge separation risk. Distributionally robust optimization, while theoretically correct, pays a steep 32% penalty for marginal safety gains. The naive two-stage SAA fails entirely under hard CPS constraints, serving as a cautionary example. We conclude that adaptive re-optimization and separation hedging are complementary rather than competing strategies, with the recommended deployment combining rolling horizons with chance-constrained separation buffers. The treatment preserves the original model's sub-minute solvability and operational constraints, offering a practical pathway to robust ALP implementation with tunable safety-punctuality trade-offs explicitly priced.

Prediction-Powered Adaptive Inference with Pretrained AI Models for Contextual Bandits

Gabriel Sargent, University of North Carolina at Chapel Hill

In adaptive experiments, statistical inference is essential for reliable decision-making and scientific discovery. Often in these settings, collecting labeled data is expensive, but decision-makers have access to large unlabeled datasets and strong pretrained AI models that can predict outcomes. Effectively leveraging these predictions in online experiments poses fundamental challenges: AI predictions may be inaccurate, and data collected under adaptive policies are inherently non-i.i.d., invalidating classical inference techniques. To address these challenges, we propose a Prediction-Powered Adaptive Inference (PPAI) estimator that integrates unlabeled data, predicted labels, and adaptively collected labeled data. We establish asymptotic normality of the PPAI estimator under mild conditions on the data-collection policy, enabling valid confidence intervals and hypothesis tests for a broad class of Z-functionals. The estimator incorporates a data-driven tuning mechanism that weights AI predictions according to their informativeness, guaranteeing that the asymptotic variance is no worse than that of the labeled-only baseline, and is strictly smaller when predictions are informative. Numerical experiments and a movie recommendation application further support the theory, illustrating efficiency gains with informative AI predictions and robust performance with inaccurate predictions.

Rank-one Convexification for Quadratic Optimization Problems with Switching Constraints

Soobin Choi, Unviersity of Southern California

We study the convexification of convex quadratic optimization problems with sign-switching constraints, which can be formulated as mixed-integer quadratic optimization problems by introducing binary variables to model the sign-switching behavior. We first derive the convex hull of the epigraph of a rank-one quadratic function subject to sign-switching constraints. Building on this rank-one convexification, we develop copositive and semidefinite relaxations for general convex quadratic functions. As an application, we derive convex formulations for support vector machines with 0--1 loss and demonstrate that these formulations produce robust estimators in the presence of anomalies and outliers.

Multi-agent Adaptive Mechanism Design

Renfei Tan, Massachusetts Institute of Technology

We study the sequential mechanism design problem in which a principal seeks to elicit truthful reports from multiple rational agents while agents’ beliefs are unknown. We introduce Distributionally Robust Adaptive Mechanism ($\dram$), a general framework combining insights from both mechanism design and online learning to jointly address truthfulness and cost-optimality. Throughout the sequential game, the mechanism would estimate agents' beliefs, then iteratively updates a distributionally robust linear program with shrinking ambiguity sets to reduce payments while preserving truthfulness. Our mechanism guarantees truthful reporting with high probability while achieving $\tilde{O}(N\sqrt{T})$ cumulative regret, and we establish a matching lower bound showing that no feasible adaptive mechanism can asymptotically do better. The framework generalizes to plug-in estimators ($\dram +$), supporting structured priors and delayed feedback. To our knowledge, this is the first adaptive mechanism under the general settings that maintains truthfulness and achieves optimal regret when incentive constraints are unknown and must be learned.

The Value and Curse of Redundancy in Designing Fork-Join Systems

Chutong Gao, Northwestern University

We consider a design problem for (n,k) fork-join systems with redundancy under fixed total service capacity. In an (n,k) system with n homogeneous servers each with rate 1/n, a job creates n tasks, one for each server, and the job departs when any k of its tasks complete. The task-size distribution incorporates variability from both the job side and the server side---For each job, its task for server i has size X S_i, where X is the job-side factor shared by all tasks from the same job, and S_i are i.i.d. server-side factors distributed as S. Under fixed total capacity 1 and task requirement k, a system manager decides the number of servers n >= k (or equivalently, the redundancy level n-k). Larger redundancy level increases the opportunity to avoid slow service-time realizations, but yields more capacity waste due to the removal of redundant tasks. Given this tradeoff, we characterize the impact of redundancy for both throughput and mean delay. For throughput, we prove the following for every positive redundancy level. (i) Value of redundancy: The stability region expands unboundedly with the server-side variability CV(S); (ii) Curse of redundancy: The critical value is strictly smaller than 1 when CV(S) is sufficiently low; (iii) Throughput-invariance case: When S is exponential, the critical value is fixed regardless of the redundancy level and the distribution of X. Numerical experiments further show that the stability region is sensitive (resp. insensitive) to the server-side variability (resp. job-side variability). Therefore, recognizing the source and magnitude of service-time variability is important when deciding whether to use redundancy, and redundancy is especially beneficial under high server-side variability. Furthermore, we prove that with i.i.d. exponential task sizes, for any positive redundancy level, the mean delay is heavy traffic optimal; whereas for the non-redundant system, the heavy traffic optimality gap is bounded by Ω(log k). Hence, we conclude that in this case, a little redundancy (n-k=1) goes a long way to reduce the mean delay.

KAMA-Ψ for CVRP: Routing Fifty Million Customers in One Minute

Aritra Banerjee, Carnegie Mellon University

We introduce KAMA-Ψ (Kernel Augmented Multiseed Algorithm over Projected Support Indexing) that operates entirely on active arc support rather than dense feasible-solution vectors. A feasible solution to the Capacitated Vehicle Routing Problem (CVRP) with n customers and r routes contains only n + r active arcs inside an arc space that grows quadratically with n; KAMA-Ψ exploits this by storing, differencing, and applying candidate improvement directions as sparse arc sets, maintained through a coordinate-incidence index that updates only directions touched by an accepted move. KAMA-Ψ is paired with a fast active-support construction pipeline–geometric giant-tour orders, word-parallel capacity-bucket dynamic programming, on-demand distance evaluation, and route-local improvement–that streams complete solutions to bounded output for independent verification, with no stage requiring a global distance matrix or dense arc vector. This construction routes 50 million customers in 59.77 seconds median (four split seeds) and 35.70 seconds mean (one seed). KAMA-Ψ also produced 20 independently validated candidate best-known-solution values–eighteen from public-BKS starts and two from complete cold starts. These results suggest active-support representation as a general template for applying Graver-style augmentation to other large-scale combinatorial problems.

Matching with Choice: Menu Design for Patient—Provider Assignment

Naveen Raman, Carnegie Mellon University

The rise in provider turnover forces health systems to frequently rematch patients and providers. High-quality patient–provider matches can improve patient experiences and lead to better downstream health outcomes by encouraging healthy behaviors. However, designing mechanisms to match patients and providers is challenging because patient preferences must be estimated from limited and noisy data, leading to substantial uncertainty at the time of matching. Moreover, matching policies are designed offline but deployed online, as patients arrive sequentially and consume limited provider capacity. To tackle this, we propose a patient--provider matching approach where we offer each patient a limited menu of providers. Such a menu hedges against preference uncertainty. The central challenge lies in properly constructing these menus. We prove that the general problem is weakly NP-hard, and establish an information-theoretic lower bound: any algorithm incurs regret that scales linearly in the noise level and in total provider capacity when the market is over-subscribed. Two natural baselines fail: offering a single provider per patient incurs regret that grows linearly in the number of patients, while offering all providers can perform arbitrarily poorly relative to the optimum. To overcome these limitations, we develop the Scenario-Averaged Marginals (SAM) algorithm, a sample-average approximation that explicitly incorporates uncertainty in patient utilities and provider capacity. We prove SAM's regret matches the lower bound up to constants when the number of samples is large. On a semi-synthetic environment calibrated to a representative provider network using Medicare data, we demonstrate that our proposed solution improves overall utility by up to 12% over commonly used baselines, while empirically finding that capacity-aware menu design can reduce geographic inequity as a byproduct. Our results demonstrate that patient choice is a powerful operational tool for mitigating preference uncertainty. Importantly, offering more options does not necessarily improve quality or access when provider capacity is constrained. Instead, carefully designed, capacity-aware menus balance patient preferences with provider availability to improve utility and reduce inequity.

Prizes & Awards

We are thrilled to announce a significant increase in this year's competition budgets! Prizes will be awarded based on evaluation by a panel of faculty judges, along with a "Fan Favorite" award voted by general attendees.

Guidelines & Logistics

• Presentation Format: Accepted participants will have approximately 5 minutes to present their poster to the faculty judges, followed by a Q&A.
• Printing Support: To help with logistics, CMU INFORMS will print accepted posters free of charge if they are submitted as a PDF at least one week prior to the event.
• Judges: Our panel of faculty judges for YinzOR 2026 will be announced soon.